Research On The Dynamic Path Planning Of Unmanned Underwater Vehicle Based On Deep Learning

The dynamic path planning ability of UUV in unknown environment is one of the important indexes to reflect its intelligence level.Traditional methods often suffer from the contradiction between environmental model accuracy and real-time planning,and in the complex environment with a large number of random motion obstacles,auxiliary strategies should be designed to get an ideal avoidance effect.These strategies are generally established under some assumptions,and often need massive and expensive equipment to support its information requirements.Therefore,it is of great theoretical and practical value to explore a simple,cheap and efficient dynamic path planning method.Depth learning is the most potential artificial intelligence algorithm.In the field of self-driving cars,the ability of dynamic planning based on deep learning has reached a practical level.However,UUV’s working environment is different from the urban traffic situation,and the sensing device is often the acoustic equipment rather than image equipment,which makes the mature convolutional neural network is not suitable for UUV’s dynamic path planning.In this paper,a more appropriate LSTM-RNN network structure is designed according to the working environment of UUV and the characteristics of sensing devices,which can be used to study the dynamic planning strategies of UUV in unknown environment.This paper mainly contains three parts:First of all,the establishment of simulation training field and teacher system based on ant colony algorithm is used to sample data acquisition network.In depth training including the construction of geometric model,model of puffing process,construction method of visualization,the basic ant colony algorithm path planning,dynamic planning strategy and any complex environment,and simulation.Firstly,a simulation training field and a teacher system based on ant colony algorithm are established to collect the training data for deep learning.This part include the methods of constructing geometric envirment model,model expansion,building visual map,and the basic ant colony path planning algorithm,as well as dynamic planning strategy and simulation verification.Secondly,deriving the basic theory of deep learning,and analyzing the characteristics and working methods of various depth network structures.We will find the advantages and problems of RNN in sequence information processing,and this is also the main reason we choose LSTM-RNN network structure for our work.Finally,we will achieve dynamic path planning based on LSTM-RNN.This part includes the derivation of LSTM-RNN’s theory,designing sample collection and sample processing methods according to the characteristics of the sonar equipmented on UUV,and designing a suitable structure for dynamic path planning.At last,the LSTM-RNN is trained and we can verfy its effection.The experimental results show that LSTM-RNN can learn the planning method of teacher system in unknown environment if enough training is done in a large enough sample set.The trained LSTM-RNN does not need an environment model nor an excessive sensing device,just one forward looking sonar can assist it dynamic planning in complex environments.